def get_lm_predictions_gpu(w,
b,
x,
y,
projection: np.ndarray = None,
device: str = 'cpu'):
network = define_network(w, b, projection_mat=projection, device=device)
accuracy = network.eval(x, y)
return accuracy
def get_top_k_lm_predictions_gpu(tokenizer,
w,
b,
x,
y,
projection: np.ndarray = None,
k=100,
device: str = 'cpu'):
network = define_network(w, b, projection_mat=projection, device=device)
distribution = network.get_probs(x, y)[0]
top_y = torch.tensor(distribution).to(device).topk(
k=k, dim=1, largest=True, sorted=True).indices.cpu().numpy()
top_words = []
for top_k_per_word in top_y:
top_k = tokenizer.convert_ids_to_tokens(top_k_per_word)
top_words.append(top_k)
return top_words
train_dir = arguments['--train_path']
dev_dir = train_dir.replace('train', 'dev')
out_dir = arguments['--out_dir']
input_dim = int(arguments['--input_dim'])
if arguments['--debias'] == 'none':
debias = np.eye(input_dim)
else:
debias = np.load(arguments['--debias'])
train_vecs, train_labels = load_data(train_dir)
dev_vecs, dev_labels = load_data(dev_dir)
x_train, y_train = flatten_tokens(train_vecs, train_labels, tokenizer)
assert len(x_train) == len(y_train), f"{len(x_train)}, {len(y_train)}"
x_test, y_test = flatten_tokens(dev_vecs, dev_labels, tokenizer)
assert len(x_test) == len(y_test), f"{len(x_test)}, {len(y_test)}"
net = define_network(word_embeddings, bias, debias, arguments['--device'])
if arguments['--debias'] == 'none':
print("Debiasing option is turned off.")
dev_acc = net.eval(x_test, y_test)
print("Dev accuracy without fine-tuning is: ", dev_acc)
net.train(x_train, y_train, x_test, y_test,
epochs=int(arguments['--n_epochs']),
save_path=f"{out_dir}/debias_ft.pt",
use_wandb=use_wandb)
def get_lm_softmax_gpu(w, b, x, y, device: str):
network = define_network(w, b, device=device)
distribution = network.get_probs(x, y)
return distribution